{"id":"bcabaa7e-3b1e-4ad0-af34-308c8631d8d9","arxiv_id":"2411.18964","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Neural operators can approximate the predictor mapping in nonlinear delay systems, and under a uniform error bound the closed loop is semiglobal practical stable.","lead":"This paper replaces the expensive numerical solver used in predictor feedback control of nonlinear delay systems with a fast neural operator, and proves the closed loop stays stable as long as the learned predictor's error is small. It matters because delay-compensating controllers are usually too slow to run in real time when the system dynamics are costly, and this work gives a formal stability guarantee for using learned approximations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 8 relies on a uniform predictor error bound along the closed-loop trajectory, but the paper only proves such a bound on a compact C1 set and never shows the neural-operator-generated control history is C1 or stays in that set.","rationale":"The reader identified the C1 regularity gap; I agree and sharpen it. The load-bearing condition for Theorem 8 is the uniform bound |\\kappa(P(t))-\\kappa(\\hat P(t))|\\leq \\alpha(\\epsilon) along the actual closed loop. The paper's route to this bound is Theorem 5, whose universal approximation guarantee is on a compact C1 domain, and Theorem 2 returns C0 outputs. No invariant-compact-set or regularity argument is supplied for the state-control trajectory generated by U(t)=\\kappa(\\hat P(t)). The proof's step p_t=p_x in Lemma 6 is exactly where C1 is used. The experiments report L2 errors and do not report the L^\\infty along-trajectory error, so they do not close the gap. This is a correctable but central gap: the theory as written does not strictly apply to the implemented controller. I therefore keep the reader's CONDITIONAL verdict: the framework is promising and likely salvageable, but the missing regularity/compactness link must be supplied.","tokens_in":21411,"tokens_out":13615,"duration_ms":133503,"concrete_test":"Analytically derive the target system (20) and Lemma 7 under the minimal regularity U\\in C^0, using the method of characteristics for the transport PDE. If the derivation requires a C1 control history (e.g., p_t=p_x fails for continuous non-differentiable U), then the implemented controller is outside Theorem 8's scope. In parallel, run the FNO controller from Section 4 on the 25 test trajectories and compute max_t |P(t)-\\hat P(t)| in L^\\infty and the numerical derivative bound of U(t); if the L^\\infty error exceeds the \\epsilon used in Theorem 8 or the derivative is unbounded, the experiments do not instantiate the theorem's hypotheses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof chain is: Theorem 5 guarantees a neural operator \\hat P with sup_{K}|P-\\hat P|<\\epsilon on a compact K\\subset R^n \\times C^1([-D,0];R^m); Theorem 8 then uses |\\kappa(P(t))-\\kappa(\\hat P(t))|\\leq \\alpha_{13}(\\epsilon) along the closed-loop solution. Two unproven links break this chain. First, the universal approximation theorem used (Theorem 2) outputs C0 functions, and the paper gives no argument that \\hat P(X(t),T_D(t)U) is C1 in its second argument or that the resulting control U(t)=\\kappa(\\hat P(t)) has a C1 history; Definition 3 and Lemma 4 are set in C1, so the predictor error bound may not apply to the very signals the controller produces. Second, even if U were C1, Theorem 5's error bound holds only on the compact set K, while Theorem 8 establishes only an L^\\infty bound on |X(t)|+\\sup|U|; boundedness in L^\\infty does not imply membership in the compact C1 set. The proof of Theorem 8 therefore uses a uniform error bound that is not shown to hold along the closed-loop trajectory. Lemma 6's assertion p_t=p_x also relies on differentiability of the control history; for C0 data the target-system derivation is not justified. This is not a minor technicality: without these links, the 'black-box predictor with uniform error bound' hypothesis is not verified for the implemented neural operator controller, so the semiglobal practical stability conclusion does not follow as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the exact predictor state in Krstic-type predictor feedback for nonlinear input-delay systems with a neural operator approximation. It defines a predictor operator on X times C1 control histories, proves its continuity and hence universal approximability by neural operators (Theorem 5), then analyzes the closed loop through a transport-PDE/backstepping representation. The main result, Theorem 8, claims B-semiglobal and epsilon-practical stability under any approximate predictor satisfying a uniform error bound. Numerical experiments on a 5-link manipulator compare FNO and DeepONet variants with successive approximations and report large speedups.","tokens_in":21668,"tokens_out":5502,"duration_ms":54645,"significance":"If Theorem 8 holds as stated, the paper would provide a useful and fairly general template: a black-box approximate predictor with a uniform error bound degrades the exact predictor-feedback guarantee only by an epsilon-practical term, while the initial-condition set can be enlarged at the cost of harder training. The cleanest contributions are Lemma 4, with its explicit Lipschitz constant, and Lemma 7, the ISS-type estimate for the perturbed transport PDE. The paper ships code and data and reports reproducible training details, which is a strength. However, the main theorem is not fully supported as written because of two load-bearing gaps: a regularity mismatch between the C1 setting of the analysis and the C0 signals produced by neural operators, and the absence of a proof that the closed-loop trajectory remains in the compact set on which the uniform approximation bound is assumed.","major_comments":[{"comment":"The stability analysis is carried out for control histories in C1([-D,0];R^m), but the implemented controller U(t)=kappa(hat P(t)) produces only continuous control histories. Definition 3 defines the predictor operator on C1 control histories, and the proof of Lemma 6 uses p_t=p_x, which requires differentiability of p and hence of the control history. Theorem 2 only guarantees a continuous neural operator, and the paper gives no argument that kappa composed with the neural operator output is C1 in time. Thus Theorem 8 does not strictly apply to the neural-operator controller used in Section 4. Please either prove the needed C1 regularity under additional assumptions on the neural operator and on kappa, or redo the transport/backstepping and ISS analysis for C0 or weak solutions.","section":"Definition 3, Lemma 6, Theorem 8"},{"comment":"The uniform approximation bound in Theorem 5 holds on a compact set K in X times C1([-D,0);U), but Theorem 8's proof uses the bound sup |kappa(P(t))-kappa(hat P(t))| <= alpha_13(epsilon) along the closed-loop trajectory without showing that the trajectory stays in K. Theorem 8 establishes only L-infinity bounds on |X(t)| and sup |U|, and L-infinity boundedness does not imply membership in the compact C1 set K. Moreover, Assumption 4 is stated on compact domains X and U, so f may not be Lipschitz outside them. The proof needs either a forward-invariance argument for K, an a priori C1 bound on the closed-loop signals, or a conditional statement that the stability estimate holds whenever the trajectory remains in the training domain.","section":"Theorem 5, Theorem 8"},{"comment":"The passage from Eq. (63) to Eq. (64) asserts the existence of beta_2 in KL and alpha_12 in K satisfying the displayed decomposition, but the composition of alpha_11 with beta_6 and with the class-K term inside the argument is only asserted, not proved. In particular, extracting a KL function from sums and compositions involving alpha_9 and alpha_11 requires a supporting lemma; class-K functions are not generally subadditive, so steps such as bounding alpha_11(a+b) need an explicit argument. Please supply the missing lemma or a detailed calculation.","section":"Appendix B.5, Eq. (63)-(64)"}],"minor_comments":[{"comment":"The final term in Lemma 7 reads sup_{0<=s<=t} |kappa(P(t))-kappa(hat P(t))|, but the integration variable inside the supremum should be s, not t; please correct to sup_{0<=s<=t} |kappa(P(s))-kappa(hat P(s))|.","section":"Lemma 7"},{"comment":"There is a missing closing parenthesis in the argument of alpha_8 in Eq. (55), which makes the displayed inequality ambiguous.","section":"Eq. (55)"},{"comment":"The notation C1([-D,0);U) is inconsistent with Definition 3, where the codomain is R^m and U denotes the control variable; please standardize the notation for the function space and for the bounded control domain.","section":"Theorem 5"},{"comment":"Please specify how the continuous-time control U(t) is reconstructed from the discrete-time neural operator evaluation (e.g., zero-order hold or interpolation) and how this reconstruction interacts with the C1 regularity assumed in Definition 3.","section":"Section 4 and Appendix C.2"},{"comment":"The paper states that detailed proofs are deferred to the extended online version; since the journal version contains a proof of Theorem 8 in Appendix B.5, the main text should either include all proofs or clearly state which theorems depend on the online appendix.","section":"Section 1 and Appendix"},{"comment":"The bottom-right panel uses a logarithmic y-axis with values down to about 1e-9; please state explicitly which norm of the prediction error is plotted and whether the error is evaluated online during the closed-loop simulation.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is actually new: instead of analyzing a specific numerical predictor, the paper proves that any black-box predictor with a uniform L∞ error bound can be dropped into predictor feedback for nonlinear delay systems and still give semiglobal practical stability. That is a clean and useful abstraction, and the perturbed transport-PDE analysis is the right way to get it. They also ship code and data, report honest speedups (roughly 30x to 420x depending on discretization), and the robotic manipulator experiments show stable tracking that matches the qualitative claim of Theorem 8. The comparison with Karafyllis and Krstic is fair, and the paper is well situated in the literature.\n\nBut there are real gaps, and the stress-test note lands on one of them. Definition 3 and Lemma 4 work with control histories in C¹, and the transport-PDE derivation (p_t = p_x) in Lemma 6 needs that differentiability. The neural operator from Theorem 2 is only guaranteed to map into C⁰, so the implemented control U(t) = κ(hat P(t)) is merely continuous. The paper never shows that the closed-loop control history is C¹, nor that the predictor operator, the backstepping transformation, or the ISS estimate in Lemma 7 remain valid for C⁰ signals. That is not a cosmetic gap—Theorem 8 as stated needs these links.\n\nThere is a second, related gap. Theorem 5 gives the uniform error bound on a compact set K, but Theorem 8 only establishes a bounded-in-L∞ bound on the closed-loop trajectory. Boundedness in L∞ does not put you back into K, so the critical bound |P(t) − hat P(t)| ≤ ε along the trajectory is not justified as written. The skipped class-K composition step from (63) to (64) is minor by comparison; that is just bookkeeping.\n\nThe experiments also report L² errors (on the order of 10⁻⁴), while the theory needs a uniform sup-norm bound. The authors do not bridge that gap, so the experimental validation is suggestive rather than tightly connected to the theorem.\n\nI do not think this is fatal. The abstraction is sound, and the gaps look fixable—either by a smoothing argument for the neural operator output or by reworking the stability analysis in a lower-regularity framework (mild solutions or absolutely continuous histories). But the proof of Theorem 8, as written, has a load-bearing missing assumption. The right outcome is a major revision that either closes the regularity gap or explicitly states the C¹ assumption and then shows the neural operator can be chosen to preserve it.\n\nWho should read this: people working in delay-compensating control and learning-based control. It deserves a serious referee; desk rejection would be wrong. My recommendation is to send it to review but with the expectation of major revision.","headline":"A genuinely useful abstraction—black-box predictors with uniform error bounds give semiglobal practical stability—but the main theorem has a real regularity gap that needs to be fixed before the result is fully rigorous.","tokens_in":22247,"tokens_out":1980,"would_cite":false,"duration_ms":19853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C23","93D15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that replacing the exact predictor in delay-compensating feedback with a learned neural operator preserves closed-loop stability whenever the approximation error is below a computable threshold.","keywords":["neural operators","predictor feedback","nonlinear delay systems","input delay","semiglobal practical stability","backstepping transformation","transport PDE","operator learning"],"falsifier":"Run the closed loop of Theorem 8 with a control signal that is continuous but not $C^1$ — for example a saturated, piecewise-linear torque — measure the largest uniform predictor error over the initial-condition set, and check whether inequality (25) holds; one trajectory exiting the predicted $\\epsilon$-ball would show the theorem does not govern the implemented neural-operator controller.","tokens_in":21129,"feed_emoji":"🤖","tokens_out":12549,"duration_ms":96530,"temperature":0.7,"pith_summary":"This paper tries to establish that the predictor—the component of delay-compensating feedback that looks one delay ahead—can be replaced by a learned neural operator without sacrificing closed-loop stability. The authors prove that if the neural operator approximates the exact predictor uniformly within an error bound, the nonlinear system with input delay remains semiglobally practical stable: trajectories converge to a ball of radius proportional to the approximation error, and the set of admissible initial states can be enlarged by training a more accurate operator. Because the proof uses only the uniform error bound, the result covers any black-box predictor, not just neural operators. The practical payoff is a delay-compensating controller whose predictor is evaluated in milliseconds rather than by solving expensive numerical ODEs, demonstrated on a five-link robotic manipulator.","feed_headline":"Neural operator predictors keep delayed nonlinear control stable","feed_subtitle":"Any learned predictor with a bounded uniform error can compensate input delays.","key_machinery":"The argument is carried by three objects. First, the predictor operator $\\mathcal P(X,U)(s)$ defined by the integral equation (9); Lemma 4 proves it is Lipschitz continuous in $(X,U)$ when $f$ is Lipschitz, which combines with the nonlocal neural operator universal approximation theorem (Theorem 2) to give Theorem 5. Second, the input delay is rewritten as a transport PDE $u_t=u_x$ with boundary $u(D,t)=\\kappa(\\hat P(t))$, and the infinite-dimensional backstepping transformation $w=u-\\kappa(p)$ maps the system to a target system whose $w$-PDE has the same transport dynamics but a boundary perturbation $\\kappa(P)-\\kappa(\\hat P)$. Third, Lemma 7 is an ISS estimate for the sup norm of this perturbed transport PDE, $$\\|w(t)\\|_{L^\\infty[0,D]} \\le $e^{{c(D-t)}}$\\|w(0)\\|_{L^\\infty[0,D]} + $e^{{cD}}$\\sup_{0\\le s\\le t}|\\kappa(P(s))-\\kappa(\\hat P(s))|,$$ which, combined with the ISS Lyapunov function for the $X$ subsystem and the inverse backstepping transformations, yields the semiglobal practical bound of Theorem 8.","core_discovery":"The paper's central claim is Theorem 8: if the neural-operator predictor $\\hat P$ approximates the exact predictor $P$ uniformly on the relevant compact set with error $\\epsilon$ below a threshold $\\epsilon^*(B)$ that depends on the chosen state/control bound $B$, then the closed-loop system under $U(t)=\\kappa(\\hat P(t))$ is $B$-semiglobal and $\\epsilon$-practical stable. Concretely, for every initial condition with $|X(0)|+\\sup_{\\theta\\in[-D,0]}|U(\\theta)|\\le\\Omega(\\epsilon,B)$, the trajectory satisfies $$|X(t)|+\\sup_{t-D\\le\\$\\theta$\\le t}|U(\\$\\theta$)|\\le\\beta_2(|X(0)|+\\sup_{-D\\le\\$\\theta$\\le0}|U(\\$\\theta$)|,t)+\\alpha_1(\\epsilon)$$ for all $t\\ge0$, with $\\epsilon^*(B)=\\alpha_1^{-1}(B)$ and $\\Omega$ defined through $\\bar\\alpha_2$. In words, any black-box predictor whose worst-case error is small enough yields a closed loop that converges to a ball of radius $\\alpha_1(\\epsilon)$, and the initial-condition set can be enlarged by increasing $B$ at the price of a harder approximation problem. At $\\epsilon=0$ the estimate recovers the stability of the exact predictor feedback.","pith_inferences":["Going beyond the paper, the main unresolved technical step is the $C^1$-versus-$C^0$ regularity gap: if the transport and backstepping identities can be shown to hold for continuous or Lipschitz control histories, then Theorem 8 would directly cover the neural-operator controller actually deployed in the experiments.","Because the theorem assumes only a uniform error bound on the predictor, the same argument should extend to learned predictors inside output-feedback or predictor-observer delay compensators, though the paper does not analyze those settings.","A likely structural limitation is compactness: Theorem 5 only guarantees approximation on compact trajectory sets, so a fully global version would probably need the approximation error to vanish with the state or an extra assumption that annihilates the boundary perturbation, as in earlier numerical-predictor results."],"forward_implications":["The stability guarantee transfers to any black-box predictor with a uniform error bound, so recurrent networks, neural ODEs, and generative models inherit the same closed-loop guarantee without new analysis.","With perfect approximation ($\\epsilon=0$) the estimate collapses to the stability bound of the exact predictor feedback design, making the neural-operator result a direct generalization rather than a separate design.","Enlarging the admissible set of initial conditions requires the uniform approximation to hold on a larger compact set, which makes the same error bound harder to achieve; the theorem makes this trade-off explicit through $\\Omega(\\epsilon,B)=\\bar\\alpha_2^{-1}(B-\\alpha_1(\\epsilon))$.","On the five-link manipulator, the FNO predictor runs about 30 times faster than one successive-approximation step and up to about 420 times faster at finer discretizations, so approximation error is traded directly for computational speedup while practical stability is preserved."],"supporting_citations":[{"why":"Supplies the exact predictor feedback stability theorem, the transport PDE representation of the delay, and the backstepping transformation that the approximate-predictor analysis extends.","marker":"Krstic (2010)"},{"why":"Provides the nonlocal neural operator universal approximation theorem (Theorem 2) used to prove the existence of an arbitrarily accurate neural operator predictor.","marker":"Lanthaler et al. (2024)"},{"why":"Supplies the numerical predictor approximation scheme used as the experimental baseline and the comparison point for the weaker assumptions of Theorem 8.","marker":"Karafyllis and Krstic (2017)"},{"why":"Gives the input-to-state stability definition and characterizations that Assumption 3 and the Lyapunov argument rely on.","marker":"Sontag and Wang (1995)"},{"why":"Provides the feedback linearization predictor for the five-link robotic manipulator used in the numerical experiments.","marker":"Bagheri et al. (2019)"}],"fun_headline_variants":["Neural operators learn delay prediction for stable nonlinear control","Semiglobal practical stability with learned neural predictors","AI predictor feedback for nonlinear delay systems, proven stable","Neural operator approximation enables practical predictor control","Fast delay compensation via neural operator predictors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that control histories are continuously differentiable, because the transport identity and the backstepping calculation are derived under that regularity, while the neural operator output used in the experiments is only continuous.","fun_headline_variants_meta":{"raw":{"variants":["Neural operators learn delay prediction for stable nonlinear control","Semiglobal practical stability with learned neural predictors","AI predictor feedback for nonlinear delay systems, proven stable","Neural operator approximation enables practical predictor control","Fast delay compensation via neural operator predictors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2284,"prompt_tokens":1000,"completion_tokens":1284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1214}},"tokens_in":616,"tokens_out":1284,"duration_ms":9298,"temperature":1.0,"reasoning_tokens":1214,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:42:30.007417+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the closed loop of Theorem 8 with a control signal that is continuous but not $C^1$ — for example a saturated, piecewise-linear torque — measure the largest uniform predictor error over the initial-condition set, and check whether inequality (25) holds; one trajectory exiting the predicted $\\epsilon$-ball would show the theorem does not govern the implemented neural-operator controller.","supporting_citations":[{"cited_title":"Input delay compensation for forward complete and strict-feedforward nonlinear systems","cited_arxiv_id":null,"evidence_quote":"Supplies the exact predictor feedback stability theorem, the transport PDE representation of the delay, and the backstepping transformation that the approximate-predictor analysis extends."},{"cited_title":"Predictor Feedback for Delay Systems: Implementations and Approximations","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical predictor approximation scheme used as the experimental baseline and the comparison point for the weaker assumptions of Theorem 8."},{"cited_title":"On characterizations of the input-to-state stability property","cited_arxiv_id":null,"evidence_quote":"Gives the input-to-state stability definition and characterizations that Assumption 3 and the Lyapunov argument rely on."},{"cited_title":"Feedback linearization based predictor for time delay control of a high- DOF robot manipulator","cited_arxiv_id":null,"evidence_quote":"Provides the feedback linearization predictor for the five-link robotic manipulator used in the numerical experiments."}],"review_version":1}